Quantification of atmospheric carbon dioxide from the Geostationary Operational Environmental Satellite (GOES East)

Technical Summary: Quantification of Atmospheric CO₂ from GOES-East

Problem Statement
Accurate characterization of atmospheric carbon dioxide (CO₂) variability is essential for understanding the global carbon cycle and verifying greenhouse gas fluxes. However, current dedicated space-based sensors (e.g., OCO-2, OCO-3, GOSAT) provide observations that are too sparse in space and time to resolve the diurnal and seasonal scales of biospheric and anthropogenic exchanges. While future missions like ESA's CO2M promise improvements, they are not expected to launch until late 2027. Consequently, there is a critical need to leverage existing instruments not originally designed for CO₂ monitoring to fill observational gaps, particularly regarding contiguous geographic coverage and high temporal frequency.

Methodology
The authors developed a physics-guided neural network to estimate dry-air column CO₂ mole fraction (XCO₂) using data from the Advanced Baseline Imager (ABI) aboard the Geostationary Operational Environmental Satellite (GOES-East). The approach integrates multispectral radiance data with meteorological and surface context through a two-stage residual deep learning framework:

1. Dataset Construction:

   • Labels: Training targets were derived from "best-quality" OCO-2 (v11.2r) and OCO-3 (v11r) Level 2 Lite soundings (2017–2023), filtered for clear-sky conditions and land-nadir geometries.
   • Inputs: The model ingests a 100-minute temporal sequence (ten 10-minute scans) of all 16 GOES-East ABI spectral bands. This is augmented with:
     • ECMWF ERA5 reanalysis data (hourly meteorological states, including boundary layer dynamics and vertical profiles).
     • MODIS surface reflectance (8-day composites) and Normalized Difference Vegetation Index (NDVI).
     • Geometric metadata (Solar Zenith Angle, Satellite Viewing Angle) used to calculate the Air Mass Factor (AMF).
   • Data Split: The 2024 dataset was held out entirely for evaluation against unseen OCO-2/OCO-3 observations.

2. Model Architecture (Two-Stage Residual Framework):

   • Stage 1 (Baseline MLP): A Multilayer Perceptron trained on predictable spatio-temporal covariates (latitude, longitude, day of year, solar zenith angle) to establish a global XCO₂ prior. This stage captures secular trends, latitudinal gradients, and seasonal cycles.
   • Stage 2 (Residual Network): A deep residual network trained to predict the deviation between observed OCO soundings and the Stage 1 baseline. Crucially, this stage is blinded to static geospatial coordinates and the year to prevent geospatial memorization. It processes dynamic inputs through three parallel branches:
     • GOES-East Temporal Branch: Processes the 16-band time series using a 1D CNN and 8-head self-attention mechanism to isolate XCO₂ signals from transient confounders (e.g., aerosols, surface reflectance changes).
     • Confounder-Mitigation Branch: Integrates ERA5 meteorology and MODIS surface data. To prevent the model from becoming a "weather-based emulator," Gaussian noise (15% perturbation) is injected into ERA5 features during training, forcing reliance on the spectral signal.
     • Physical AMF Branch: Explicitly calculates and injects Air Mass Factor components to normalize for geometric distortions inherent in geostationary viewing angles.

3. Training Strategy:

   • The model was optimized using Huber Loss to handle outliers.
   • OCO-3 data (non-sun-synchronous) was upweighted by 2x relative to OCO-2 to ensure the model generalizes across the full diurnal cycle.
   • Automated Machine Learning (AutoML) techniques, utilizing Large Language Models and Monte Carlo Tree Search, were employed to fine-tune hyperparameters and architectural choices within strict physical constraints.

Key Results

• Performance: When evaluated against held-out 2024 OCO-2 and OCO-3 observations, the full GOES model achieved a Root Mean Square Error (RMSE) of 1.42 ppm and an R² of 0.60. This represents a significant improvement over the Stage 1 baseline model alone (RMSE 1.86 ppm, R² 0.31), confirming that the model learns physical signals rather than merely interpolating historical patterns.
• Ablation Studies: Removing the GOES-East time series caused the largest degradation in performance, confirming it as the primary information source. Interestingly, removing the specific CO₂-sensitive bands (Band 5 and 16) had only a marginal impact, suggesting redundancy in the broadband spectral channels and the model's ability to synthesize complementary signals from other bands and auxiliary data.
• Case Studies:
  • Urban: The model successfully visualized XCO₂ enhancements over Los Angeles and Chicago with 10-minute temporal resolution, capturing plume structures consistent with coincident OCO-2 and OCO-3 Snapshot Area Mode (SAM) observations, as well as NO₂ data from TROPOMI.
  • Agricultural: Over the US Corn Belt, the model captured a regional-scale 2 ppm diurnal drawdown during the peak growing season (July), a variability pattern missed by sun-synchronous instruments that only overpass once per day.
• Latitudinal Gradients: The model correctly reproduced realistic latitudinal gradients and interannual variability for 2024, distinct from the training years (2017–2023).

Significance and Claims
The paper claims that while the precision of GOES-East derived XCO₂ cannot rival that of dedicated hyperspectral instruments (e.g., OCO-2's ~0.85 ppm precision), the unprecedented combination of contiguous geographic coverage, 10-minute temporal frequency, and multi-year records offers a unique capability to observe aspects of atmospheric CO₂ variability currently unseen from space.

The authors emphasize that this approach is an augmentation, not a substitution, for dedicated missions. The high spatial and temporal density of geostationary data allows for the observation of diurnal cycles and rapid transport dynamics that are invisible to current polar-orbiting constellations. Furthermore, the study demonstrates that existing instruments, when coupled with modern deep learning and auxiliary data, contain substantial recoverable information about atmospheric CO₂. The authors also note that the model inherits biases present in the OCO-2/OCO-3 training data (e.g., river network artifacts), suggesting that dense GOES observations could eventually help elucidate and correct these biases in the reference datasets themselves.